AI Virtual Workshop Series - AI for Business Strategy & Decision-Making (2 of 4)

0:03 Okay, good morning everyone. Thank you for our part two of our uh AI workshop. Today we're talking about business strategies for AI adoption. And I want to thank Pam Clark from Clark Communications for sponsoring our series. And I'm going to introduce you to Mia who is the co-founder and CTO of Venture Factory AI. Mia Kennady is a data scientist, entrepreneur, and early stage investor with a track record of co-founding tech companies in India and the US. She's the co-founder and CEO of Venture Factory AI, a global AI venture studio and also leads Beyond Z LLC, a US-based digital transformation firm. As joint manager

0:49 managing director at Venture Way, she drives global tech investments in startup ecosystem growth. An active angel investor, she serves on the board of governors at Sparks Angel Network. Mia holds advanced degrees and certifications in AI and data science from Sunni Buffalo, MIT, Stanford and Harvard. She is the founder of Chaya Foundation which supports education for underprivileged children in India and is establishing the Gai Foundation in Geneva to promote global empowerment through AI. So thank you Mia and I'll turn it over to you. Thank you so much Margie. It's pleasure to be here. Great to see all the familiar faces from last time and some new faces as well. Um, so

1:37 let me get started by sharing my screen and we'll Oh, and let me say one other thing. If anyone has questions, please put them in the chat so Mia can address them after the presentation. Thank you. You see my slides? Yes. Great. So, this is workshop two in the three-part series. We're going to talk a little bit about uh business strategies for AI adoption um and how AI can help um you know formulize your business strategy, formulate your business strategy. In today's agenda, we have AI's role in business strategy. We're going to talk a little bit about the difference between

2:24 traditional analytics and big data analytics. Um we're going to talk about predictive analytics for smarter decision making. how can you drive uh smarter decisions for your business using predictive analytics and then AIdriven market analysis and competitive intelligence. Um and then we'll go into a couple of case studies uh from companies that have been using AI uh to great benefit. So diving right in AI's role in business strategy. So what AI can do especially the uh language models the chat bots that we have what they can allow you to do is quickly brainstorm business ideas. So normally when you have a brainstorming

3:10 session you're throwing ideas and you're getting feedback uh from your team members uh who are bringing their experience uh into that discussion. And what a language model allows you to do is draw on the collective experience that's contained on the web in digital books in research papers etc. Essentially the experience that is contained in everything that that AI model is trained on. So you know we've seen um you know in sci-fi movies we see these assistants that are you know throwing out ideas and helping refine ideas. We are sort of on the cusp of that reality today. Um and what this allows you to do is quickly iterate through variations of strategies to consider feasibility. And then if you

3:57 talk to um you know Chad GPT um versus other uh maybe Gemini, Meta AI, you talk to different language models, you get a different perspective as well. Um and I I like to stress the fact that um language models are essentially just predicting the next word but they are so good at it. The training is so good that um they do give out useful information. How innovative are AI models? So what is seen is that because everything that an AI model uh uh gives out is a function of its training, they are excellent at combining ideas from different domains. So they are excellent at suggesting or

4:44 thinking through new ideas that are based on work that has been done in different fields. And in entrepreneurship we do see uh that a lot of innovation happens at the intersection of different fields, different domains. What AI models are not great at doing uh is coming up with things that have never been done before. If you can imagine um the invention of electricity before it was a reality. If you can imagine the suggestion of the internet before it uh existed, those are things that had never been seen by the world before. Um and those are the kind of ideas that AI models are not great at coming up with. Uh but because they have a wealth of training data, they are great at combining uh ideas from

5:30 different domains. And very often in entrepreneurship, that is how new ideas are arrived at. How can AI help in business strategy? So one is that it can give you datadriven insights. Um there is a lot of value in in instinct and intuition. Um and you can back that up with datadriven insight from your data. It can uncover drivers of profit and growth that are existing in your current business. Uh it can analyze help you to analyze the effectiveness of current approaches. Um you can monitor your business real time which is very valuable. AI can also help you ask the right questions just by having a better view a viewpoint into

6:17 your data. Um it helps you to ask the right questions and thereby re uh arrive uh at interesting um strategy points. How do you spot AI opportunities in the business? Um so one is to look at what are the outcomes uh that would impact your business the most. Um it is always better to start from the why. Uh why would you approach a certain solution? Why would you want to take on a certain um problem? And then figure out how and then you look at what are the relevant data points. Very often for AI currently repetitive tasks that have well- definfined outputs are great candidates for AI because then you have a wealth of training data and you can use that to

7:03 train the AI models. Another thing to look at when you're looking at uh what may be ripe for AI is what is the risk involved in incorrect predictions. Um you can imagine that in healthcare um if you're going to predict a certain diagnosis then you wouldn't wouldn't want to put it out there without um a doctor having seen it as well because this are high risk. Um but in terms of Netflix for example recommending a certain movie that's low risk. So um you want to think of um if you're starting out your journey, you would perhaps want to consider lower risk initiatives first um and then uh work your way up to higher risk um initiatives in your business. So this is looking at AI impact by

7:50 industry as a percentage uh of the total impact that analytics can bring in. And you can see right up there is retail um especially because of e-commerce. There's a lot of data in um how customers engage uh with content, what customers buy, what customers are interested in uh and it's ripe for analytics. So retail is an industry that is will be impacted a lot by AI methodologies. Transport and logistics is another one and I mentioned it uh in workshop one as well. Um how AI can be used um you know to uh prevent disruptions in the supply chain uh to improve logistics um and it does impact cost in a big way. Automotive and

8:35 assembly is another area. uh predictive maintenance we spoke about in manufacturing that's one there's uh robotic process also automation so a lot of uh impact that can come in that field as well banking high techch industries um are also going to see a lot of impact in AI because they are already in the tech space and they have a lot of data and they have the tools um to implement AI so we're going to talk a little bit about what differentiates It's traditional analytics and big data analytics. Um and we'll talk about um what is big data to begin with, right? Essentially, it is data that exceeds the capacity or capability of conventional

9:21 methods. We're talking about terabytes of data, huge data sets that we're collecting. Now, the reality today is that almost all businesses are collecting data um and they have a wealth of data at their disposal. So big data necessitates storage to accommodate. So you do need to look at what you how are you going to store this data. You need tools to visualize and analyze this data. uh because when you're talking about highdimensional data sets there could be um you know hundreds of dimensions um you cannot visualize it easily and there are tools that help you narrow down the dimensions uh to into the features that are most relevant so that you're seeing the data that affects your outcome the most. It's very important to visualize

10:06 to be able to visualize data because that is where um a lot of insight comes from and you also of course need uh the software the tools to process u big data. What is the difference in an in analysis um from traditional analytics versus big data analytics? Traditionally uh we have a question to begin with um and then you have a hypothesis uh for example in in uh drug discovery you're you're saying will this drug help this um condition and then you have a hypothesis you're saying yes it will and then you go and collect the data you run a clinical trial um and you have a control group

10:53 and a test group um and then you receive you arrive at an answer uh by collecting your data. So the quest the the process starts with the question and then with a hypothesis. It's very hypothesis based um and you collect data based on the hypothesis that you're trying uh to clarify and then you get an answer. However in big data analytics in today's age of data uh it is driven by the data. So you start with your data, you run exploration on it, visualization on it, um different types of analytical approaches and you see what that data throws up. What are the correlations in the data? Um what is the what are the things that you may be able to predict?

11:39 What are some interesting uh statistics on that data? And from there you draw actionable insight. Um, so you don't have a a condition that you're starting with or a question that you start with, but you're starting with your data and you let your data drive um action. So big data allows you to ask new questions. It's when you explore your data that you realize uh that some questions can be um answered with that data. Um it asks new questions. It may allow you to answer existing questions in a better way. uh because of correlations or you might see clustering similarities between users that you didn't realize existed. Um you might see

12:25 underlying um you know correlations that you can use um in your business in your approach to business. So when you're analyzing data when you have big data and you're um doing analysis on it the workload um surprisingly is um a big part of the workload is in prepare preparing that data right first you would acquire the data you as in as in you're collecting data from your customers or from your business and you have to make sure that it's correct you have to do some error checking so that is number one acquisition and verification that takes up some time but the am most amount of time and cost takes up is taken up in preparing the data right you need to

13:12 make it make the data in a format that uh AI models can be you can use for training and that is what takes up the bug so even when you're budgeting for um you know for your data you need to keep in mind that data preparation takes up a lot of time take ups a lot of u cost as well analysis takes up time and effort and then communic communicating right you need to be able to communicate the insights that you're getting across the organization um to the people that need to make the decisions that need to make the changes um so roughly this is how um the workload is split up what is so once you have your data and you're going to train an AI model for your business what is the life cycle of

13:57 an ML project of a machine learning project or an AI project um so step one if you going from the top down. Uh you would plan the project. Uh you would choose the metrics that you're going to decide whether the project is uh successful or not. Uh you would need some baselines as in if you're going to use an AI methodology to do a certain task. It's necessary to have a baseline in place to evaluate how your system is doing. Very often human performance is a good baseline uh for tasks that humans uh do. So that is your baseline and then you're going to see how uh your AI mount model performs and then you need the code base etc. Once you do the setup you have to do you know your strategy you have to collect the data and label the

14:44 data u or prepare the data right labeling comes into place when you're doing supervised learning but there are also others unsupervised learning methodologies. So that's the next step, the data step. And then you're going to train your model, debug the model. Um, so you would need to make a model choice. You need to train it, you need to debug it, you need to improve the results. So there's a lot of iteration there. And then finally, you would deploy the model and you would do testing on it. Uh, but you can see these right uh arrows on the side that are going from the boxes up. Um and that highlights that model uh an ML project is very iterative in nature, right? It is when you um you know get to perhaps

15:31 the ingestion stage of data, collecting your data, you may realize this particular type of data is not possible to collect. Then you need to go back to the drawing board and redefine your problem to be able to use other kinds of data. or um you know you may collect your data, you may start training and debugging and realize that um you know you need to label your data in a different way for the model to be uh to perform well. So then you go back uh to preparing your data and you make those changes. Um again after deploying you may realize um that the metrics you chose for um success don't translate uh well in a real world setting or you may

16:16 realize that you know the actual data that you're seeing in model usage is a little bit different from the data that the model is seeing at training um and then you go back to the drawing board. So ML project development is very iterative and you need to um factor that in when you pro uh when you plan when you budget uh time um and cost. It's very iterative in nature. There's a lot of going back and forth um before you hit that sweet spot and before it works well. How do you go from um analytics to decision making? How do you actually impact your business? So, that's what we're going to look at. Smarter decisions. AI allows you to make smarter

17:04 decisions because um it gives you insights from your data. So, it's forecasting the future based on what it has seen historically. Um so, you can plan based on your data. datadriven planning also better risk management because you can see what disruptions may be coming see what the market may be looking like um what you need to be prepared for so those are the kinds of things that AI can help you do better but how do you go from predictions to better decisions right so wherever there is correlations in your data uh you can make good predictions right um but correlation and causation are different things. Um just to highlight that for

17:49 example um and uh you know there is it was found that students who have music who are gifted in music um do better academically right um but then given music is it is does giving them music lessons uh does giving a child music lessons make them better academically? So is music a causative factor for academic performance or is there an other underlying uh factor? So it was found that um you know there are people who are gifted um who do well at music and at academics but studying music doesn't necessarily transfer translate to better academic performance. So that's the difference between

18:34 correlation and causation. So whenever you have correlations in your data, you can use that to make um accurate predictions, right? But you need to understand causation to be able to take a decision. So how do you go from a prediction that a model is making to the right decision? For example, you use your sales data um and you say that a certain product a user is likely to buy if you make a certain change. There is a correlation there. That is what the model is predicting. But what is the right decision to take? To understand that effect to understand cause the only way to go about it is to

19:19 act. So you need to um compare action versus no action. So essentially you need to um have hold out sets right you run a sort of pilot program on a part of your users a part of your users base you don't give them the others uh that same effect and you see how the two perform example in the drug trial example you give some people uh the actual drug you give the hold out uh the control group uh a placebo and you compare the results right so to to clearly understand the impact of a decision. You do need to take um take action and compare that uh action versus no action. So there is an in inevitable cost involved in that

20:05 process because uh one is you need to run the experiment and two assuming that um the drug for example in our example assuming that the drug does give benefit then there is a hold out set a control group who didn't get the drug um or in in a in a in a consumer space if there is a product that will sell well uh and people are buying there is a hold out set who are not exposed to that product. So there is a small cost involved. Uh but that is uh worth worth doing. The experiment is worth doing. So how do you experiment with different options? Assuming that your AI model has predicted a certain course of action, how do you compare action versus no

20:51 action? One standard way of doing it is AB testing where um you have one group who is exposed to um a certain outcome and one group that is not and you compare the results. A slightly different one is multi-arm bandit and contextual bandit algorithms which are also where you expose different groups to different scenarios but based on incremental results that are coming in you modify uh for example if in a drug control um example if you see that the drug is beneficial then you move more people into the test group you expose more people to the um to the drug. This is not done in drug testing typically but in uh in uh non-farmaceutical

21:38 non-drug space where you're talking about consumer goods um multiam bandits are often used. Um e-commerce websites use them all the time where some of the users are recommended certain product and if they see that users are buying then more users are um exposed to it before waiting for the complete results of this test to come out. so that you the cost of not exposing some users to a product that they're likely to buy is reduced. This is some more uh details on uh these uh algorithms but um I'm going to get into it only if maybe if there are questions in the end I'm happy to get into the details of it.

22:23 So what drives prediction accuracy when you're building an AI model? Uh when you're training an AI model for your business, what drives prediction accuracy? One is data quality. Uh AI models very much are garbage in garbage out systems. So if you have data that is not accurate or data that is erroneous, the model is not going to um be able to predict accurately. So data quality is one, data quantity is another. Uh based on the uh number of parameters that the model has, uh it will need uh a fair amount of data to be able to um really be trained well. Model choice is another that drives prediction accuracy. Different types of problems require

23:09 different types of models. There are models that work well for uh certain types of scenarios. So making the right model choice is important. And then domain expertise um really understanding what are the relevant features for a certain problem uh what are the number of features that you need to be able to predict a proper outcome. Um so domain expertise has a lot to do if you think about uh for example um diagnosing X-ray images that is an area where AI models have shown a lot of success but you do need a doctor to first come in and say these are the things that we look at in an X-ray to be able to diagnose what the diseases are. Similarly in in your business you will know what are the things that are relevant for a certain outcome and that is very much a part of

23:56 the puzzle. So when often given um budgeting constraints and when you're at a point where you have trained a model and you now want to improve performance um which is better would you invest more in data or would you invest more in in another model better algorithms or better data which is more important? uh definitely goes without saying both are important but very often high quality data has the most impact. So if you had to make a decision between getting better data and uh experimenting with another model uh get better data if if there is a constraint on that.

24:42 So now we're going into AIdriven market analysis and competitive intelligence. How can AI solutions help in that regard? Competitive intelligence. So AI helps you to understand what your competitors are doing at scale, right? You can monitor online data to understand competitor pricing. Um you can monitor uh online data to understand what products the competitor has out there. Um you can analyze and summarize market data, right? You can use public financial data and customer feedback um to benchmark competitive performance. Um you can um you know analyze the

25:29 sentiments that um reviews have to your customers products. Um and the thing with AI is that it helps you do this at scale. You know you can have a a human uh go and look in online, look at a few websites, look at a few reviews. Um but AI allows you to do it at scale you know really mine all the data that is out there and draw insights from it which is very hard to do with human resources. So it is the scale that AI brings in and the insights from that scale that really help. Nike is um has been using AI for um competitive intelligence to understand what competitor pricing is, what models, shoe models the competitors are coming out with. So really AI helps

26:16 you to understand your competitors at scale very efficiently. Market analysis for your own products. Um, again you can mine data from social media, from news, from e-commerce platforms to really look at what your customers are talking about, what are they looking for and what are they likely to want in the coming months. Um, what products or services are likely to trend and when. Um the anatomy of a trend uh has changed because of this age of data and this a social media because earlier uh to it took more time for things to become popular and for products to trend and then trend lasted

27:02 trends lasted longer. But today because because of how connected people are because of social media uh trends come in and go much faster. Um so it's great to use AI to understand what are the going what are the upcoming trends what should you be ready for um and you can also analyze the nature of those trends to see are they going to last um should I really ramp up for this one uh should I sit this one out so you can really analyze your market at scale again it is the scale that changes the game because you're able to really analyze data from a vast array of sources is um and draw insights from that to be able to prepare

27:47 yourself as a business for what the landscape is going to look like product market fit. This is again something that AI can help you analyze um very deeply. One is um so I've got this image execution capabilities is you know what your business is currently capable of doing and AI can help you understand better what your core business strengths are and abilities are. What are you good at doing? You can really analyze the data to see what you've been doing well. What are you good at doing? What are your strengths? The other circle is the potential products, right? um what are the products that you could offer given your strengths? What are the products that

28:33 you could offer the market? Um again AI can help you uh ideulate and strategize on that front. The third circle is market needs, right? What are customers looking for? What is going to sell? Um what are the likelihoods that if you put a product out there, someone's going to buy it? Yeah. Again, uh like we just spoke, AI can help you mine online data to really understand what people are looking for and where the execution capabilities, your potential product portfolio and the market needs meet. That is your true product market fit. And using AI and um methodologies, you can draw more insight into each of these areas to really analyze product uh

29:20 market fit so that you know uh from a strategic perspective where you want to take your business. And now we come into uh a few case studies uh about companies that have been using AI um to great benefit. The first one is Clara. I mentioned Clara last time as well. Um and we're going to get into a bit more detail today. U Clara is again a financial services company. It's a buy now pay later service provider. Uh they produced uh they provide payment processing for the e-commerce platforms. What learner did is it integrated generative AI very early on into the shopping experience into customer service um into for their

30:07 employees right so they have a customer service chatbot which we sort of touched upon last time um reduced weight times higher volumes of calls serviced um that's what they've been seeing so Kiki is their AI assistant um it's powered by open AI What Kiki allowed them to do is multilingual customer service and the customer service calls um essentially can be stratified into simple calls, more complex calls and so on. And the very simple ones, it makes a lot of sense to have an AI take care of for you. Someone calls and and says, "Have you received my payment?" An AI chatbot can take care of that. um and

30:54 really reduce the workload on the human workforce. Um this AI assistant also is trained to pro give information to foster healthier financial habits. It gives realtime updates about balances and payment schedules. So what Clara found that 2/3 of their entire customer service calls about 1.3 million conversations per month are now handled by this AI agent. um about around last year uh the CEO said that it is doing the equivalent of 700 full-time human agents available in 23 markets available 24/7 communicates in more than 35 languages so those advantages are hard to beat

31:40 connected data so in this journey um the customer service agent is one of the uh AI solutions that they use but more importantly they have made an organizational change. Uh they have connected data from across the organization. Last time we spoke about the centralized data repository bringing data from across the organization into a central uh repository and that is what CLA has done. So they've removed their siloed data architecture and it has allowed them to link uh people to projects, documents to teams, processes to outcomes and that is where the real insights um are available when you're able to see across um uh across domains

32:27 across departments um to really analyze what is affecting your outcomes. companywide AI adoption, right? They took on an um uh AI, generative AI and made it available across the organization um and they helped their employees to be able to use it. Uh they gave their employees the freedom um to be able to use and get you get used to using AI tools. um you do need to enable a bit of that uh to also address the fear that is um there around using AI. Uh so you need to be able to let your workforce experiment um and get a feel for it. Um and what they've seen is

33:13 improved productivity. They've seen professional growth. Uh they have their communications, marketing and legal team using these generative tools um very well and a lot. So I mentioned they they you know um there was a $40 million profit in 2024 first quarter of 2025 they've seen 15% revenue growth um and they've uh attributed a lot of it to AI right uh so clearly AI is making great impact but um this is from 5 days ago so it says CLA reverses on AI says customers like talking to people, right? Where do we place this in our AI landscape? And so

34:02 remember I talked about iteration earlier, right? A lot of AI is about putting it out there, looking at how it's doing and iterating, right? Um, and this headline was a bit sensational, but that is not the whole truth. There's been further interviews with the CEO where he clarified that Clara is still going full steam on AI. It's just that in their customer service sector they have realized that there are some calls that do need human agents. So last year CLA paused their hiring um and it made waves. They're saying you know it's going to replace the human workforce. Today their stance is that it's augmenting the human workforce. They still do need humans. Um but then AI

34:48 takes care of a lot of calls as well. So the their business has grown um and that has been attributed to AI but there are still a lot of things that they do need humans for. So for those um businesses that are considering using AI for customer service one key um find one key uh insight is that it needs to be easy for a person calling in to reach a human agent. Right? So anything that the AI can handle it should handle but the human the person calling in needs to be know that anytime they want to they can reach out to a human and that makes a lot of difference right um so to throw light on the overall business this is also from I think 3 4 days ago where

35:36 they're saying their revenue per employee so to nearly $1 million thing thanks to AI efficiency push so um here you see that it's not they're not reversing on AI they is still very bullish on AI but they are learning as they go and I want to draw back your attention to the image about the ML project life cycle where it's all about iterating you do need to go back and um you know iterate till you find the sweet spot so very often you decide to use AI you've gathered and prepared the data you've trained your model you've deployed it and you have great confidence in it but sometimes you have underwhelming results. But the key is to know that, you know, that's not the end

36:21 of the journey. Um, and that's not where you want to say, you know, AI is not going to work for me, but rather you deploy, you monitor, you do user testing, you modify as needed, you maybe you need to get more data, maybe you need to do model training, you know, you do what is needed, you deploy again. Um, and that's where uh you have impact. So I feel I think there are two loops that on that are ongoing in AI development. One is a very fast loop. You need to iterate and uh improve your model very fast and once it gets to a certain level of performance you deploy it. But there's also a slower loop once deployment happens where then you're looking at how is it you know performing

37:07 in the real world. Um and you're also again iterating and modifying there. That's a slightly slower loop. Um so in the initial development there's a very fast iterative loop that uh should go on and once you deploy there is still also a lot of uh modifying um and learning that needs to go back into the um into the product development before you get that real impact in the business. So the next um uh case studies from BMW um AI for strategy right BMW has been using analytics and AI in a very big way um so they've been using predictive AI to analyze their global sales to see what are the local trends to see

37:55 consumer behavior and to also tailor their products and features. uh they they have a lot of inventory as you can imagine um and they've been using AI to uh optimize their inventory. So they have superior product market fit. Uh they have reduced overstocking and others they've used generative AI to to really empower their workforce and to improve efficiency. Right? They've used um they have a group AI assistant that their employees can use. They've also empowered their employees to build their own apps using generative AI. Um so they are a platform company. They have um a lot of data that they collect a lot of they have their own data sets. Um they have in-house AI capability but they've

38:43 also empowered the workforce to look at scenarios where they think generative AI tools can help them and then to build apps um that will do that very quickly. So that boosts productivity that fosters innovation. Um and in this vein um there is also um you it's important to think about when you should build uh AI in-house and when you should buy it's not you don't always need to build all the solutions inhouse you don't always need to outsource it but it's good to use a combination of both operationally also the insights that they the AI systems that they have drawn from their data have uh allowed them to really streamline their operations

39:28 So, and they have a strict very strict data life cycle management as you can imagine they collect um a lot of data um and you can't really save all the data that you collect uh because that would be uh from a budget perspective that would be formidable. So, you really do need to think about from the data that you're collecting uh in your business what do you really want to save? Um and sometimes that changes right you you save some data today another use case comes up tomorrow you may need to save some other data you may need to start collecting other data and that's okay so it's an iterative process but it's very often not feasible to save everything that you're collecting but the idea is to have a close eye on it um so that you

40:16 do save what you need you throw what you don't need and when that changes when uh the use case changes you change your uh data strategy accordingly. Also, life cycle management. You've been saving some data. Um but you need to look at who is using that data, what use case is that data fueling, and what impact does that data have on your business? And if there's, you know, data that you've collected that you're not using at all over a a given length of time or or it's not driving any use cases that have real impact in your business, then you really want to look at uh stopping to stopping saving that data perhaps um you know in favor of other data that could be more useful for you. Um so here BMW also links use cases

41:05 to data assets so that they know what use cases each data set is driving. They're able to analyze and measure um uh the value of a data set and also they've prepared their landscape for AI agents. I mentioned AI agents last time um where you know you have language models that have access to external tools and this access works through APIs. uh which is program uh you know a sort of programming interface. Um so the uh applications that you have in your um in your organization need to be um need to have those APIs uh that the language model can use. So you do need to manage your application landscape in a way that

41:52 you can uh have AI agents right and um random um rag which is retrieval augmented generation that is a very popular uh methodology today we talked about knowledge base management last time where you have language models that are answering questions based on documents in your organization um that is an example of uh retrieval augmented generation But for their incar assistance, um BMW found that um the time the latency that rag systems have doesn't work when you're driving a car and you have a question, you need an answer immediately. So they're looking at uh continuous pre-training of models. So they're doing a lot of things. They have uh built the framework that allows them

42:39 to consider different solutions to iterate fast. Um they collect data. They have very strict data governance. um and that has been driving their organizational strategy. Some early wins that BMW had uh one is data centralization. Again, it's a a theme that pops up uh in most companies that have been using AI. Um so they consolidated their vast data into a data hub and they did it started doing it as early as 2017. They have very strong data governance. They have data stewards who are responsible for the quality, the security and data compliance. Um, and they have a business-driven strategy because they are saving data, they are doing analytics on it, and they've made it possible for their workforce to um,

43:27 assess access those analysis. Um, they've become a business uh, that is driven by data. Right? So the impact uh just to highlight some of the key results again um improved product market fit um and reduced inventory. So they're able to understand what their customers want um what they need to be prepared for um so they're able to reduce inventory. They're able to put out better products um that their customers really want. uh enhanced customer satisfaction um right because uh their customer service is now powered by data powered by AI they're able to do u better customer service um which actually fosters uh brand loyalty

44:15 because customers feel that they are um you know their needs are being met uh increased operational efficiencies as I mentioned and accelerated innovation so the very the streamlining as they go. Um so they're very running a very tight ship. Uh they monitor their data very closely. Um so I have another um one more example which I thought I'll decide to get into if time permits. Uh Mona is a company I think that is uh close to you guys based out of um Netherland. So Moderna is also a company that integrated AI and analytics into their business early on. Um again they have a integrated digital

45:00 library of data right so no silos all their data is um in one place also top-down strategic intent um this is also very important the biggest hurdle to AI adoption is often not technology but culture so it's very important uh to really communicate um top down the strategic intent around AI adoption in your business to um to show up its value early on um and to allow employees to engage um and to build trust in the system. So Mona also has a very strong top- down strategic intent uh that is communicated uh across the company. They're cloud native uh

45:46 meaning that uh they have a lot of data on the cloud and that's easily scalable and AI is integrated into their core business strategy across the board. So from customer service to product development, knowledge management to legal um and so on. They've spent uh a lot of time and e effort in employee upskilling. They have a an AI academy where employees go regularly to um you know get upskilled. They're empowering employees. Um they have easy to use tools for quick integration and um the employees have the freedom to build custom GPTs um generative models that they think will really benefit their daily tasks. They have the freedom to build those custom

46:32 GPTs and that's something we saw uh with CLA as well. So they have also a lot of strategic partnerships with other players in the industry. So Mona uh partnered with OpenAI early on. Um they have customuilt AI chatbots. Uh they gave their employees chat GPT enterprise access um so that they could build custom GPTs. So in the two months in around 2 months they had 750 custom GPTs in use. That was as of April last year. as of today they have more than 3,000 custom GPTs in use. Um so that is the sort of um uh freedom that the employees are given and they are seeing um a lot of productivity from it. Uh they had a pilot GPT pilot

47:17 called dose ID early on uh which used to review and analyze clinical data to see what dosage might work uh in terms of vaccines. It's able to visualize large data sets um so it really impacts um analysis. Another partnership that uh Moderna has with is AWS um Amazon Web Services and they have this OC3 omni channel cloud contact center is a solution that AW has AWS has uh that allows you to do across theboard customer service um and also analyze your customer service. It really um look allows you to offer the ideal customer journey through the organization. Uh better experiences for all su all stakeholders not just

48:04 customers but also other um uh providers um policy makers uh for everyone. So the OC3 is a channel that they've been using from AWS. Um it allows personalized services. Um, so yeah, that is it from me for today. Uh, thank you. I'm I'm looking forward to the questions. Let's see. Thank you so much, Mia. Um, you can put your questions in the chat or you can um just ask Mia. We have a small group here. Um, so I know a lot of that was over my head, Mia,

48:52 but that's me. Um, I do I do have a question. How do people create their own chat box? I mean, where do you start to create your own chatbot? So, you have to start with existing conversations. Um, right? If you have a customer service uh arm already um then you sort of record those conversations and you use that to train the chatbot. Um so if you see like a chat GPT for instance it is conversational already right you can ask a question and have a human-like response already but then you want it to know your business. So that is where you would fine-tune the model uh to know uh

49:40 what is relevant for your business. What are the kinds of conversations that happen in your business. Another one is is domain specific terms. So there may be uh you know abbreviations and terms that are very specific to your sector to your industry uh that the chatbot may not be uh aware of. So those are the kinds of things that uh get done in finetuning right and you can create uh custom GPTs um that that you give access to uh some documents so that the GPT is familiar with terminology from your business. I do like what you said about CLA. How many times have we all got on trying to get a hold of someone and you're stuck

50:27 in the thing and you're yelling customer service operator? It's so frustrating for people. So, I like that these companies are using, you know, um this technology, but at the same time, they do have to keep the customer the customer service. Yes. Okay. Here's a question in the chat. Two real world case studies are supposed to be discussed in today's session. Are we done with these cases or I had missed them? Well, we talked about CLA, BMW and Moderna. So in in uh you know in addition to what you were saying about CLA um very much uh at the heart of innovation is the

51:14 human element. Um right. So you do need to keep that uh very squarely in the picture. Um and you know in um in customer service there are a lot of tasks that are very simple and uh customers do appreciate uh expediency in those but you just need to quickly you have a quick question you want to know about your account. Um and often when it's a human operator then there is that time for them to access your records and look into your this thing and give you the answer. Uh but an AI agent can do it much faster. So those kinds of uh responses AI does very well. Um but there may be more complicated scenarios where you do want a human agent. Um, and

52:00 you know, it's it's the the idea is to make that access easy. Well, I know next week we're going to be talking about AI for marketing and sales growth. So, I know a lot of people are very interested in that. Yeah. So, I look forward to that. Um, Pam has Do you want to ask it yourself, Pam, or do you want me to just read it out of the chat? Hang on here. You mentioned using AI to assess what your company does well and where improvements could be could be made. Um,

52:45 how would you recommend doing that? So it comes from collect the data that you collect on your customers. Um you know you have for example you have marketing material. Um you also measuring uh the number of people who are um engaging with that material. You make a change in your marketing material. Uh and then you look at how customers are engaging with that. What is the click-through rate for example? or if it's a sales platform, what is the purchase uh you know, how many customers are purchasing? So from analyzing the response to your company's um offering, you're able to understand what you do well, what you do better, how can you

53:30 improve. Um and like I mentioned, the AB testing, the multiam bandits, those kinds of algorithms allow you to run simultaneously two three versions of um your service offering to understand um which might be better for your market. Okay. Thank you. Rebecca, any questions? Not today. I'm It's all great information. Thank you. Thank you, Becky. Any I have one question, Mia. Um when Sean um started our AI

54:18 uh journey a few weeks ago, he talked about when you're in GPT, you you should probably have a private account and not just uh no one had ever explained that to us before. So there were many people in the room that were um using GPT but not understanding that we probably should have an account. What is your thought about that? So we can use Margie as an example. If she went in and and said, "Okay, I'd like uh you to give me an analysis of every chamber in Montgomery County or in Maryland, and I want to know the revenues they bring in. I want to know the membership. I want to know

55:04 how long they've been around. I want to know who the who's running it. Every question she can possibly think to ask. Should she do this um by not logging into an account or should she have an account? And why is it important to have an account? Okay, so um one of the great biggest advantages of having an account um is that the language model then has your context, right? So then uh because it keeps um a memory of your conversations um and internally it has a knowledge graph uh of you know what you've talked about and the connections between those knowledge pieces. Uh so that is one reason to have an account in terms of

55:51 privacy. um everything that you send to the language model like a chat GPT that is running uh not uh private to your organization but as a public service uh is then used for further training that model um so that that comes into the public sphere in that sense. So for the example that you stated where uh for example Margie were to ask about all the other chambers in the industry and um you know what to look at there is no privacy concern there at the outset because most likely it is going to use publicly available data about the other chambers. Uh but in terms of unique analysis, you may you may have an idea

56:36 uh which then you want to use public data to think through but that you may not want that idea uh at a very early stage uh to be referenced uh publicly right but also if you're having conversations about your health for instance you may ask you know these are my parameters um you know what is something about your health or something that's private to you um you may not want to send that uh you know to a language model. So there was a case where um a question was asked to ch and it responded and in that response there was some snippets from a New York Times article um but that article was actually

57:24 behind a pay wall right but then because someone asked at some point would have used in that training and it went into the training data it then became public information. So there are various ways in which um data that you send to the language model then comes into the public sphere, right? I guess you just have to think about that when you're asking questions and make sure that if you don't want to publicize broadly and give this information out, you just need to keep it close, right? And not not do it. Yeah. Okay. Okay. And organizationally, people that deal with um a lot of sensitive data may then

58:11 choose to have a locally hosted language model rather than sending it to a um you know a chat GPT like model through API. Okay, good. Thank you. Sure. Okay, everyone. Thank you very much and we'll see you uh next Friday on May 30th for AI for marketing and sales growth. So I wish everyone a happy Memorial Day weekend. Um Mia, have a safe trip coming back and we'll see everyone next week. Thank you so much. Thanks everyone. Bye bye.